LAUSR

Detecting and localizing anomalous behavior in the surveillance video is explored and spatiotemporal model, which jointly learns the appearance and motion-based feature is proposed. The general solution is to learn from the normal-data as reference models and uses various hands designed features. However, huge variations can occur within normal-behavior patterns. It is a challenge to represent higher-level concepts of a normal or abnormal event explicitly from raw input data. In the proposed framework, spatiotemporal features learned at various hidden layer are analyzed. Based on the learned representation, the reconstruction of video volumes are performed. Finally, the structural distortion based abnormality score is computed by considering luminance, contrast, and structural information to detect the presence of abnormality and localize them. Further, we also explored the performance of GMM and one-class SVM in a given scenario. The proposed structural distortion based abnormality detection and localization are evaluated on the publicly available UCSD and UMN dataset. The performance of the developed system is found to outperform the existing state-of-art methods for detecting and localizing abnormality at the frame as well as pixel-level. Recently, deep architecture is also found to be vulnerable to adversarial attacks and can easily be tricked to fool the system. However, most of the existing attacks are designed for the classification task. In this work, we utilize the gradient-based approach to generate adversarial samples for an abnormality detection system. Finally, we build the defense mechanism to detect the abnormality in the presence of such adversarial attacks.